Spokes.wiki Search About
Software Source Code source ↗ source url updated Tue Jun 23 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

DeerFlow — source summary

An open-source “long-horizon SuperAgent” harness from ByteDance (MIT; ~73.2k stars; Python on LangChain/LangGraph). A full harness instance that runs autonomous agents on tasks spanning minutes to hours — research, coding, report/slide/media generation — rather than single-turn chat. Delivered via Telegram, ingested 2026-06-23. v2.0 is a complete rewrite sharing no code with the original “deep research” framework.

What it bundles

The wiki’s recurring harness layers, shipped together by a big-tech vendor:

  • Sandboxes — isolated Docker/local execution where agents read/write files without touching the host.
  • Memory — persistent, user-controlled long-term memory (preferences, style, domain knowledge) across sessions; the agent-memory substrate.
  • Tools & Skills — web search, files, bash, custom tools via MCP servers, and progressive-loading markdown-defined skills for research/reports/slides/media — i.e. the agent-skills pattern again, inside a vendor harness.
  • Sub-agents — a lead agent spawns scoped child agents that run in parallel and synthesize back: the orchestrator → fan-out mechanism as a built-in.
  • Message gateway — reach the agent from Telegram, Slack, Feishu/Lark, WeChat, WeCom, DingTalk without a public endpoint.

Why it’s interesting here

  • The harness-vendor stratum gains a big-tech entrant. Alongside microsoft (Foundry / Agent Framework), Google (adk), and langchain, ByteDance now ships a production agent harness — another data point that “structure around the model” is consolidating into vendor platforms, not just community tools. It’s built on LangChain/LangGraph rather than replacing them — a harness layered on the middleware substrate langchain supplies.
  • Convergence on the same parts. Sandboxes + persistent memory + markdown skills + parallel sub-agents + long-horizon loops is the exact composition the wiki keeps seeing (gstack, zouroboros, microsoft-scout, hermes-agent); DeerFlow is the same attractor reached independently at ByteDance scale, with markdown skills (agentskills-spec-shaped) and a long-horizon loop.
  • Reflexive note. The message gateway (agents driven from Telegram/Slack/Feishu) is the same chat-as-control-surface this hub uses — a productized version of the Telegram-driven agent.

Caveats

First-party repo (T2): feature list and “production-grade” framing are self-reported; no neutral benchmark of its long-horizon completion vs. simpler harnesses. Dated snapshot (v2.0).

agentic-coding-harness · agent-orchestration · agent-memory · agent-skills · durable-agents · langchain · bytedance · gstack · microsoft

Related pages

Links to

Linked from